paper-reach

name	paper-reach
version	0.1.0
description	Run evidence-aware literature screening with Paper-Reach. Search candidate papers, screen abstracts conservatively, review full text when available, and emit structured JSON.
argument-hint	screen topic query.json, review local PDFs, run literature screening for a research question
allowed-tools	Bash, Read, Write
homepage	https://github.com/Dai0-2/paper_reach
repository	https://github.com/Dai0-2/paper_reach
author	Paper-Reach contributors
license	MIT
user-invocable	true

Paper-Reach Skill

Use this skill when the user wants a rigorous literature-review workflow rather than ad hoc web search.

Purpose

Paper-Reach is a reusable screening workflow for:

• candidate paper retrieval
• conservative abstract screening
• optional full-text fetching and review
• evidence extraction
• rubric-based ranking
• JSON-first outputs for literature review and gap analysis

Host-Agnostic Workflow

1. Find the Paper-Reach installation root if you need repository files.
2. Read AGENTS.md for repository-wide expectations.
3. Use the relevant sub-skill under skills/:
   • skills/paper-search/SKILL.md
   • skills/paper-reader/SKILL.md
   • skills/paper-ranker/SKILL.md
4. Prefer the installed paper-reach CLI over reimplementing screening logic in prompts.
5. Consume the generated JSON and reason from explicit evidence and conservative decisions.

Finding The Repository Root

If the current working directory is not obviously the repo root, search these common locations and use the first directory that contains both AGENTS.md and paper_reach/cli.py:

for dir in \
  "." \
  "${CLAUDE_PLUGIN_ROOT:-}" \
  "${GEMINI_EXTENSION_DIR:-}" \
  "$HOME/.claude/skills/paper-reach" \
  "$HOME/.agents/skills/paper-reach" \
  "$HOME/.codex/skills/paper-reach"; do
  [ -n "$dir" ] && [ -f "$dir/AGENTS.md" ] && [ -f "$dir/paper_reach/cli.py" ] && PAPER_REACH_ROOT="$dir" && break
done

If you only need execution and the package is already installed, you can skip repository path discovery and call the CLI directly.

Execution Preference

Prefer these commands:

paper-reach doctor
paper-reach example-query
paper-reach screen --input query.json --output screen.json --high-recall --retrieval-limit 120
paper-reach review --input query.json --output review.json --local-path ./papers
paper-reach fetch-fulltext --input query.json --output review.json --download-dir ./downloads
paper-reach run --input query.json --output result.json

Fallback if the console script is unavailable:

python -m paper_reach.cli doctor
python -m paper_reach.cli screen --input query.json --output screen.json

Guardrails

• Never treat title-only relevance as enough for final selection.
• Treat abstract-only support as coarse evidence.
• Keep unsupported papers as ambiguous or need_fulltext.
• Prefer JSON outputs over free-form prose when handing results back to the user or another agent step.
• If remote retrieval fails, fall back to offline mode or local files instead of inventing evidence.